marcodsn/iol-lfm-baseline

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Jul 6, 2026License:lfm1.0Architecture:Transformer Featherless Exclusive Cold

LFM2.5-1.2B-Instruct is a 1.2 billion parameter instruction-tuned causal language model developed by Liquid AI, built on the LFM2 architecture. It is designed for efficient on-device deployment, offering fast inference speeds on CPUs and NPUs with a low memory footprint. This model excels in agentic tasks, data extraction, and RAG, supporting a 32,768-token context length and multilingual capabilities across English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.

Loading preview...

LFM2.5-1.2B-Instruct: On-Device AI

LFM2.5-1.2B-Instruct is a 1.2 billion parameter instruction-tuned model from Liquid AI, part of the LFM2.5 family of hybrid models. It is specifically engineered for on-device deployment, delivering high performance in a compact size. The model boasts extended pre-training on 28 trillion tokens and large-scale multi-stage reinforcement learning, enabling it to rival much larger models in quality while maintaining efficiency.

Key Capabilities & Features

  • Optimized for Edge Inference: Achieves 239 tok/s decode on AMD CPU and 82 tok/s on mobile NPU, operating under 1GB of memory.
  • Broad Compatibility: Supports llama.cpp, MLX, and vLLM from day one, with available quantized formats (GGUF, ONNX, MLX) for diverse hardware.
  • Multilingual Support: Handles English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
  • Tool Use: Features robust function calling capabilities, allowing for integration with external tools and agentic workflows.
  • Long Context: Utilizes a 32,768-token context length for processing extensive inputs.

Ideal Use Cases

LFM2.5-1.2B-Instruct is particularly well-suited for:

  • Agentic tasks
  • Data extraction
  • Retrieval Augmented Generation (RAG)

It is not recommended for knowledge-intensive tasks or programming. Benchmarks show it outperforms other sub-2B models like Qwen3-1.7B and Gemma 3 1B IT across various intelligence and instruction-following metrics, making it a strong choice for efficient, high-quality on-device AI applications.

Finetunes
1 models